Transcribing Audio and Video via API: Under $0.50 for a 45-Minute Podcast

Speech-to-Text sounds simple — until you actually try to build the pipeline yourself. Downloading media files, stripping audio tracks from video, slicing long files that exceed size limits, sending them in chunks, then stitching the timestamps back together. If you're starting with a pile of links — podcast episodes, meeting recordings, lecture videos — getting to a clean transcript often means writing a lot of glue code.

A Speech to Text tool deployed on the Apify platform aims to bundle all of these steps into one: one link in, one transcript out. This article breaks down its capabilities, cost structure, and where it fits best.

One Link In, One Transcript Out

This Speech to Text tool supports MP3, M4A, WAV, FLAC, OGG, MP4, MOV, WEBM, and more — any link or uploaded file with audio can be processed. It handles the entire pipeline end-to-end: downloading, audio extraction, chunking, transcription via Whisper large v3, and timestamp realignment. A three-hour recording requires one input and produces one output.

Each output is fairly comprehensive:

• Clean transcribed text with timestamped segments
• Ready-to-use SRT or WebVTT subtitles
• Retrieval chunks with timestamps for RAG pipelines
• Automatic language detection across 90+ languages
• Optional speaker labels

The API call itself is straightforward — a single curl command is enough to trigger a run:

curl -X POST "https://api.apify.com/v2/acts/lergassy~speech-to-text/run-sync-get-dataset-items?token=YOUR_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"urls":["https://example.com/episode-42.mp3"]}'

The real value here is that it fully abstracts away engineering complexity. For developers who just want text output and don't want to maintain chunking and reassembly logic, what gets saved isn't just a few lines of code — it's an entire round of debugging edge cases.

---

Whisper large v3 is an open-source speech recognition model released by OpenAI in late 2023. It's the largest and most accurate model in the Whisper family (approximately 1.5 billion parameters), excelling particularly at multilingual recognition and accented speech, with support for transcription and translation across 90+ languages. Because the model weights are fully open, the community has produced numerous inference acceleration solutions — Groq being one of them. Through its proprietary LPU (Language Processing Unit) chips, Groq delivers several times the inference speed, significantly lowering the per-minute transcription cost. This explains why the "bring your own Groq key" option cuts the unit price from $0.009 to $0.004: Groq's hardware advantage makes the same model run faster and cheaper.

---

Cost Breakdown: ~$0.41 for a 45-Minute Podcast

The tool's pricing model is based on a combination of minutes processed and number of files:

• Using the built-in transcription key: $0.009 per minute, no additional signup required
• Bring your own Groq key: $0.004 per minute, with Groq costs billed directly to you at fractions of a cent per minute
• Per file processed: $0.002
• Speaker labels: an additional $0.006/minute when enabled; no charge when disabled

Using this structure, a 45-minute podcast episode costs approximately $0.41 with the built-in key (45 × $0.009 + $0.002); with a Groq key, that drops to around $0.18. That's where the "under $0.50" headline comes from.

One hidden cost worth noting: for video platforms like YouTube, TikTok, Instagram, X, and Vimeo, audio must be fetched through residential exit nodes, which adds $0.09 per link. Direct file links incur no such fee.

The author also offers a practical tip: for YouTube videos that already have captions, using a dedicated subtitle-scraping tool is far cheaper — there's no reason to run speech recognition on content that already exists as text. This is a cost optimization worth keeping in mind.

---

A residential exit node refers to a proxy network that routes requests through real home broadband IP addresses. Platforms like YouTube and TikTok actively detect and block bulk requests from data center IPs, but residential IPs behave like ordinary users and are far more effective at bypassing these restrictions. Acquiring and maintaining residential proxy infrastructure is significantly more expensive than standard server egress — which is why processing video platform links carries an extra $0.09 charge. That fee is essentially the cost of the proxy network, not the transcription itself. Direct file links (e.g., CDN-hosted MP3s) don't need to circumvent platform anti-scraping measures, so they incur no such cost.

---

Use Cases and Limitations

Based on the use cases the author outlines, this type of tool is best suited for: podcast show notes and site search, meeting and interview archiving, turning entire course libraries into searchable knowledge bases, and generating subtitles for your own videos. These scenarios share a common profile: many links, low per-item processing value, but significant total volume — exactly the kind of workload that benefits from batch API consumption.

The limitations are equally clear. It can only transcribe what is audible — this is a universal limitation of Whisper-based tools: heavy background music or crosstalk will degrade accuracy. There's also a hard constraint around media accessibility: content behind a login or in private storage can't be downloaded; you'd need to provide direct file links or upload files manually.

In short, this tool solves the "tedious pipeline" problem, not the "better recognition" problem. The underlying model is still Whisper large v3 — the quality ceiling is no different from deploying Whisper yourself. The difference lies in the engineering wrapper and the convenience of pay-as-you-go pricing.

---

Retrieval chunks are an output format designed specifically for Retrieval-Augmented Generation (RAG) pipelines. The core idea behind RAG is to split long documents into small metadata-tagged segments stored in a vector database, then retrieve the most relevant chunks at query time and inject them into the prompt so the language model can answer based on the source material. When transcript output preserves timestamps for each segment, retrieval results can be pinpointed to exact moments in the original audio — especially valuable for podcast Q&A, course search, and similar use cases where users want not just relevant content, but the ability to jump directly to that moment. SRT and WebVTT are two mainstream subtitle file formats: SRT has the broadest compatibility, supported by virtually all video platforms and players; WebVTT is the native HTML5 standard, ideal for embedding in web-based players.

---

The Real Value of Tools Like This

Placing this tool in broader context, it represents an increasingly common product pattern: taking a mature open-source model (Whisper), wrapping it in engineering infrastructure, and lowering the barrier to entry through a standardized API and transparent pricing. For teams that occasionally need transcription, or want to embed transcription into a larger RAG pipeline, this kind of "link in, text out" black box is often the easier path.

That said, it's worth staying clear-eyed about the trade-off. At very high volumes, self-hosting Whisper inference or using the Groq API directly is usually more cost-effective. The author even acknowledges that content with existing captions should take a cheaper route. This tool is best positioned for users with moderate data volumes who don't want to invest engineering resources in maintaining their own pipeline.

(Disclosure: The author noted that their Apify links include a referral code.)